app.py

import os
import torch
from torch.utils.data import Dataset, DataLoader
import pandas as pd
import numpy as np
import joblib
import gradio as gr
import plotly.graph_objects as go
from io import BytesIO
from PIL import Image
from torchvision import transforms, models
from sklearn.preprocessing import LabelEncoder, MinMaxScaler
from gradio import Interface, Image, Label, HTML
from huggingface_hub import snapshot_download
import torch_xla.utils.serialization as xser
import s2sphere
import folium

token = os.environ.get("token")

local_dir = snapshot_download(
    repo_id="robocan/GeoG_23k",
    repo_type="model",
    local_dir="SVD",
    token=token
)

device = 'cpu'
le = LabelEncoder()
le = joblib.load("SVD/le.gz")
len_classes = len(le.classes_) + 1

class ModelPre(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.embedding = torch.nn.Sequential(
            *list(models.convnext_small(weights=models.ConvNeXt_Small_Weights.IMAGENET1K_V1).children())[:-1],
            torch.nn.Flatten(),
            torch.nn.Linear(in_features=768, out_features=1024),
            torch.nn.ReLU(),
            torch.nn.Linear(in_features=1024, out_features=1024),
            torch.nn.ReLU(),
            torch.nn.Linear(in_features=1024, out_features=len_classes),
        )

    def forward(self, data):
        return self.embedding(data)
    
# Load the pretrained model 
model = ModelPre()

model_w = xser.load("SVD/GeoG.pth")
model.load_state_dict(model_w['model'])

cmp = transforms.Compose([
    transforms.ToTensor(),
    transforms.Resize(size=(224, 224), antialias=True),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])

def predict(input_img):
    with torch.inference_mode():
        img = cmp(input_img).unsqueeze(0)
        res = model(img.to(device))
        probabilities = torch.softmax(res, dim=1).cpu().numpy().flatten()
        top_10_indices = np.argsort(probabilities)[-10:][::-1]
        top_10_probabilities = probabilities[top_10_indices]
        top_10_predictions = le.inverse_transform(top_10_indices)
        
        results = {top_10_predictions[i]: float(top_10_probabilities[i]) for i in range(10)}
        return results, top_10_predictions

# Function to get S2 cell polygon
def get_s2_cell_polygon(cell_id):
    cell = s2sphere.Cell(s2sphere.CellId(cell_id))
    vertices = []
    for i in range(4):
        vertex = s2sphere.LatLng.from_point(cell.get_vertex(i))
        vertices.append((vertex.lat().degrees, vertex.lng().degrees))
    vertices.append(vertices[0])  # Close the polygon
    return vertices

def create_map_figure(predictions, cell_ids, selected_index=None):
    fig = go.Figure()

    # Assign colors based on rank
    colors = ['rgba(0, 255, 0, 0.2)'] * 3 + ['rgba(255, 255, 0, 0.2)'] * 7
    zoom_level = 1  # Default zoom level
    center_lat = None
    center_lon = None

    for rank, cell_id in enumerate(cell_ids):
        cell_id = int(float(cell_id))
        polygon = get_s2_cell_polygon(cell_id)
        lats, lons = zip(*polygon)
        color = colors[rank]

        # Draw S2 cell polygon
        fig.add_trace(go.Scattermapbox(
            lat=lats,
            lon=lons,
            mode='lines',
            fill='toself',
            fillcolor=color,
            line=dict(color='blue'),
            name=f'Prediction {rank + 1}',
        ))

        # Adjust zoom level if selected prediction is found
        if selected_index is not None and rank == selected_index:
            zoom_level = 10  # Adjust the zoom level to your liking
            center_lat = np.mean(lats)
            center_lon = np.mean(lons)

    # Update map layout
    fig.update_layout(
        mapbox_style="open-street-map",
        hovermode='closest',
        mapbox=dict(
            bearing=0,
            center=go.layout.mapbox.Center(
                lat=center_lat if center_lat else np.mean(lats),
                lon=center_lon if center_lon else np.mean(lons)
            ),
            pitch=0,
            zoom=zoom_level  # Zoom in based on selection
        ),
    )

    return fig


# Create label output function
def create_label_output(predictions):
    results, cell_ids = predictions
    fig = create_map_figure(results, cell_ids)
    return fig

def predict_and_plot(input_img, selected_prediction):
    predictions = predict(input_img)

    # Convert dropdown selection into an index (Prediction 1 corresponds to index 0, etc.)
    if selected_prediction is not None:
        selected_index = int(selected_prediction.split()[-1]) - 1  # Extract index from "Prediction X"
    else:
        selected_index = None  # No selection, default view

    return create_map_figure(predictions, predictions[1], selected_index=selected_index)




# Gradio app definition
with gr.Blocks() as gradio_app:
    with gr.Column():
        input_image = gr.Image(label="Upload an Image", type="pil")
        selected_prediction = gr.Dropdown(choices=[f"Prediction {i+1}" for i in range(10)], label="Select Prediction to Zoom", value=None)
        output_map = gr.Plot(label="Predicted Location on Map")
        btn_predict = gr.Button("Predict")

        # Update click function to include selected prediction
        btn_predict.click(predict_and_plot, inputs=[input_image, selected_prediction], outputs=output_map)

        examples = ["GB.PNG", "IT.PNG", "NL.PNG", "NZ.PNG"]
        gr.Examples(examples=examples, inputs=input_image)
    gradio_app.launch()